Cactus Compute has released Whistle, a 16.9 MB speech-to-text model designed to run on a CPU while supporting seven languages, word timestamps, and probabilities. A recent tutorial and benchmark evaluation tested the compact model against OpenAI's Whisper tiny and base models on identical LibriSpeech clips using an Apple M4 processor.
The evaluation found that Whistle holds a Word Error Rate lead over greedy Whisper base configurations on test-other samples, though it shows no significant performance gap when compared against beam search. Furthermore, the model demonstrates a notable speed advantage that gradually narrows as audio clip lengths increase, while also providing features like keyword biasing, embeddings, and speech-to-tool calls.
Why it matters: This evaluation highlights the viability of highly compact, CPU-friendly speech models for resource-constrained environments.